chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[AMD] Fix accuracy while using --enable-dp-attention (#19247)

Browse Source 
Co-authored-by: yichiche@amd.com <jacky.cheng>
This commit is contained in:
jacky.cheng
2026-02-25 12:50:28 +08:00
committed by GitHub
parent ab0f608788
commit e138f7960a
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/layers/attention/aiter_backend.py Normal file → Executable file
View File

@@ -203,7 +203,10 @@ class AiterAttnBackend(AttentionBackend):
# need to fall back to non-persist
# only use mla_ps_kernel when fp8 kv_cache
# for non-fp8 kv_cache on tp8, use non-persist kernel to avoid performance degradation
if self.num_head == 16 and self.kv_cache_dtype is not fp8_dtype:
# head_num=16 (tp8 perf issue), head_num=128 (unsupported, like tp1 or --enable-dp-attention with tp8-dp8)
if (
self.num_head == 16 or self.num_head == 128
) and self.kv_cache_dtype is not fp8_dtype:
_use_mla_ps_kernel = False
fast_mode = False
intra_batch_mode = False